NVIDIA's high-density AI product is a rack, not a server in a random chassis. This hub does two things: it reads what NVIDIA has actually published about that cabinet, then maps the Taiwan-listed ODMs that bolt it together. Power, water, wide-bandgap switches and the CoWoS package under the GPU are separate topics — 800 VDC, liquid cooling, SiC / GaN and CoWoS. Every company below has an English memo page with briefings of its earnings calls.
Part 1 — What NVIDIA actually specified
Source: NVIDIA's GB200 NVL72 product page, the MGX rack programme, and the 800 VDC whitepaper 800 VDC Architecture for Next-Generation AI Infrastructure (Jared Huntington and Mike Tu).
The product fact
GB200 NVL72 is one copper domain: 72 Blackwell GPUs and 36 Grace CPUs, plus NVLink switch trays, in a liquid-cooled rack. Successors — GB300, Vera Rubin, then Kyber at 576 Rubin Ultra GPUs and 1 MW+ — stay on that path. Air-cooling a tray at that density is not the reference design.
MGX is the modular rack NVIDIA sells as a partner programme. The 800 VDC architecture's first phase is an MGX-compatible side power rack in 2H 2026. Full 800 VDC production lines up with Kyber in 2027.
Who NVIDIA named — and who it did not
The published 800 VDC ecosystem list names Taiwan power-component houses (Delta, Lite-On, BizLink, Lead Wealth) and Richtek on silicon. It does not name a Taiwan ODM. Rack assembly is a different fact from a named power SKU.
Foreign brokers have named Hon Hai / Foxconn among the Taiwan partners building Kyber. Quanta and Wiwynn sit in the same slot by what they tell their own shareholders, not by an NVIDIA vendor PDF.
The conversion the ODM has to absorb
Specifics worth holding onto:
- The ODM is the integrator. NVIDIA specifies the domain. The listed company that ships the cabinet is the one answering for yield, liquid loop, busbar, test power and a factory that can take 1 MW-class racks.
- Consignment versus buy-and-sell is how that BOM hits the P&L. High-ASP GPUs inflate revenue and crush gross-margin percent unless the customer ships the chip. Every name on this page is negotiating that mix.
- 800 V is the 2027 cabinet. 2026 calls date Vera Rubin / GB shipments. Do not read a 2026 ship date as an 800 VDC design win.
Part 2 — The Taiwan companies
Read the slot before the thesis. Assembling NVL72 for a hyperscaler is a different fact from building a sovereign-AI Level-12 rack, and both are different from making the sheet-metal chassis.
Rack-scale ODMs
These names sell the finished cabinet. Dual-tagged here and on the 800V hub, because the same rack is what 800 VDC is written against.
Hon Hai / Foxconn (2317.TW) — brokers' Kyber build partner. Cloud and networking crossed 51% of Q2 2026 sales. Self-made content is above 50%, chips excepted: busbars, CDUs, cold plates, manifolds, QDs, plus in-house SiC. Vera Rubin is guided to production readiness in Q3 2026 and shipments in Q4, with a 50% share target; that is the 2026 cabinet, not Kyber. ASIC share is aimed at 40% or more. Capacity: 2,000 racks a week in 2026. None of the 2026 calls named 800 V. Six notes, through Q2 2026.
Quanta Computer (2382.TW) — notebook ODM that is now a server company by mix. Servers were above 80% of sales from Q1 2026; AI was 75–80% of server revenue in the first half, guided to 80% in the second. Q2 2026 revenue NT$1.04tn. AI-server revenue is guided to double in 2026; order visibility is talked about through 2028. Capacity is guided to double by end-2026 versus end-2025, and to double again in Taiwan when the acquired AUO Huaya plant starts in 2028. The Huaya building already has high-voltage supply for outgoing test — factory test power, not NVIDIA's 800 VDC rail. The 13 August 2026 call never said 800 V. One note, Q2 2026.
Wiwynn (6669.TW) — Wistron-group CSP specialist. 2025 AI mix crossed 50%, of which ASIC was about 90% and GPU the rest. GB200 GPU projects printed in Q4 2025 and then rolled off; new NVIDIA and AMD GPU platforms plus new ASIC platforms are guided to 2H 2026. Level-12 (full-rack) work is described as tenders and sovereign AI, taken selectively. Air- and water-cooling capacity are both in; factory power is planned to 2028. The 26 February 2026 call never said 800 V. A later AGM (25 May 2026) did mention HVDC, CPO and cooling as ecosystem work — that sentence is not in this note. Q2 2026 was a board results release, not a 法說會. One note, FY 2025 / Q4.
Adjacent — on the 800V hub, not dual-tagged here
Chenming (3013.TW) — chassis and racks, including liquid-cooling lines. It studies both the OCP ±400 V and NVIDIA 0–800 V camps. It is sheet metal, not a hyperscaler ODM book.
Blocked for now. Inventec (2356) and Wistron (3231) assemble servers, including AI. They are not written here until a completed dump shows the NVIDIA rack as the primary story rather than a notebook-and-server mix. Wiwynn's May 2026 Yuanta and Morgan Stanley conference decks are broker days, not company earnings calls.
How to read these memos
Each memo is an English briefing of one Taiwan earnings call. Figures are as stated by management and are not independently audited. The Chinese transcript and the Chinese memo behind each one stay off this site; the English article is the published work.
Where a company's most recent earnings call in our archive is old, the memo says so rather than implying the position is current. "AI server" mix percentages are company-stated and are not comparable across issuers — Foxconn's cloud-and-networking line, Quanta's AI-within-servers ratio and Wiwynn's AI-versus-general split are three different P&L cuts.